One language, top to bottom
The framework, the inference network, the storage protocol, and the chain are all Rust. There is no Python interpreter in the serving path.
05Deep learning framework
RUMUS
Deep learning without the Python runtime.
Summary
A deep learning framework written natively in Rust with a WGPU compute backend. It is the training and execution layer underneath the models OmniNode distributes, which keeps the whole inference path in one language and one memory model.
Role in the stack
Removes the last foreign runtime from the stack. Rust from the node binary to the kernel.
Portable compute
A WGPU backend targets Vulkan, Metal, and DirectX from the same kernels, which is what lets a consumer laptop join OmniNode.
Publicly held to account
Open source, past one hundred stars, and the substrate for the company research program.
Architecture
- Graph Rust, no interpreter
- Kernels WGPU compute shaders
- Device Vulkan, Metal, DirectX 12
- OmniNode Shards the resulting model
Specification
Values are taken from the protocol repository and from live chain parameters.
RUMUS specification
| Language | Rust, native |
|---|---|
| Compute backend | WGPU |
| Target APIs | Vulkan, Metal, DirectX 12 |
| Python runtime | None |
| Role in stack | Trains and executes the models OmniNode shards |
| Distribution | Open source |
Endpoints and source
- rumus.suminnovation.xyz app
- Source repo